Directory
Every model below is a public repository. Filter the list in this browser. With the script off, the full list is still here.
44 models
No model matches these filters.
Qwen's 7B instruct model for ordinary chat. It is the size a single desktop GPU is often asked to hold.
Smallest pack we can size: Q4 estimate · 5.2 GB, including the allowance.
An 8B-class Qwen text model. The directory lists it for chat and step-by-step answers on hardware you own.
Smallest pack we can size: Q4 estimate · 6.0 GB, including the allowance.
A small Qwen model that takes text and images. Useful when the machine is a laptop rather than a server.
Smallest pack we can size: Q4 estimate · 5.6 GB, including the allowance.
A mid-size Qwen model for text and images. Larger than the 4B sibling, still in desktop range on a 4-bit estimate.
Smallest pack we can size: Q4 estimate · 8.8 GB, including the allowance.
Qwen's 27B-class vision-and-text model. The 4-bit estimate is the pack to read first. The full-weight repo is a different, larger download.
Smallest pack we can size: Q4 estimate · 19.8 GB, including the allowance.
Qwen's 7B instruct model for questions about images. It is a vision model, so the fit row adds memory for the image tower.
Smallest pack we can size: Q4 estimate · 8.1 GB, including the allowance.
An 8B Qwen instruct model for image and text inputs. The publisher repository is the source for which files to download.
Smallest pack we can size: Q4 estimate · 8.3 GB, including the allowance.
A 32B Qwen instruct model. Most 12 GB cards will not hold the 4-bit estimate. A 24 GB card is the first size to check.
Smallest pack we can size: Q4 estimate · 20.5 GB, including the allowance.
A 72B Qwen instruct model. Plan on a rented GPU unless the machine has workstation memory.
Smallest pack we can size: Q4 estimate · 44.0 GB, including the allowance.
A 14B reasoning distill published by DeepSeek on a Qwen base. Read the license on the repository before you ship it in a product.
Smallest pack we can size: Q4 estimate · 9.6 GB, including the allowance.
DeepSeek V3.2, a large text model. The parameter count is the whole model. A desktop fit is unlikely. The rented-GPU row is the one that matters.
Smallest pack we can size: Q4 estimate · 380.9 GB, including the allowance.
DeepSeek's V4 Flash checkpoint dated in the repository name. Treat the size on the public record as the size. Do not assume a small active-parameter figure unless that record states one.
Smallest pack we can size: Q4 estimate · 171.3 GB, including the allowance.
DeepSeek V4.1 Flash. The public record classes it as image-and-text. The fit estimate uses that larger allowance.
Smallest pack we can size: Q4 estimate · 425.8 GB, including the allowance.
Z.ai's GLM-4.5 Air text model. Air in the name is the publisher's label. The gigabytes below are still computed from the published parameter count.
Smallest pack we can size: Q4 estimate · 64.8 GB, including the allowance.
Z.ai's GLM-5.3 Flash. The public record classes the repository as image-and-text. Check the license on the card before commercial use.
Smallest pack we can size: Q4 estimate · 182.7 GB, including the allowance.
Microsoft's Phi-4 text model. It sits between a laptop model and a 32B model. Read the 4-bit row before the full-weight row.
Smallest pack we can size: Q4 estimate · 9.6 GB, including the allowance.
The smaller Phi-4 instruct model from Microsoft. Built for machines that cannot hold the full Phi-4 weights.
Smallest pack we can size: Q4 estimate · 3.1 GB, including the allowance.
Google's Gemma 3 4B instruct model. The weights are under Gemma's own terms, which the license row names. Accept those terms on the publisher's site before downloading.
Smallest pack we can size: Q4 estimate · 5.4 GB, including the allowance.
A Gemma 4 instruct model. The name marks a mixture-of-experts shape. The parameter count shown is the total on the public record, not a guess at how many are active.
Smallest pack we can size: Q4 estimate · 18.7 GB, including the allowance.
Google's Gemma 4 31B instruct model. A 4-bit estimate is the first number to compare with your card. The full repository is larger.
Smallest pack we can size: Q4 estimate · 21.7 GB, including the allowance.
Meta's Llama 3.1 8B instruct model. Meta gates the download. The fit estimate does not change the fact that you have to accept their terms.
Smallest pack we can size: Q4 estimate · 5.9 GB, including the allowance.
Meta's Llama 3.3 70B instruct model. This is a rented-GPU model on the 4-bit estimate unless the machine is a workstation. The download is gated by Meta.
Smallest pack we can size: Q4 estimate · 42.8 GB, including the allowance.
A 1.7B instruct model from Hugging Face. Small enough that the interesting question is quality for the task, not whether the weights fit.
Smallest pack we can size: Q4 estimate · 1.9 GB, including the allowance.
OpenAI's open-weight gpt-oss model at the 20B size. The repository is the source for the license and the files. This desk does not resell access to it.
Smallest pack we can size: Q4 estimate · 14.0 GB, including the allowance.
The 120B gpt-oss model published by OpenAI. Read it as a rented-GPU model unless the fit row says otherwise for a specific pack.
Smallest pack we can size: Q4 estimate · 68.3 GB, including the allowance.
Mistral's 7B instruct model, v0.3. A known desktop size. Confirm the license on the card. It is not Apache on every Mistral release.
Smallest pack we can size: Q4 estimate · 5.0 GB, including the allowance.
Mistral Small 3.2 at 24B, instruct weights dated in the repository name. Compare the 4-bit estimate with a 16 GB or 24 GB card.
Smallest pack we can size: Q4 estimate · 15.7 GB, including the allowance.
Prism ML's ternary Bonsai 2 pack, about a 27B-class model stored as GGUF. A normal 4-bit estimate would misstate these packed weights, so this page uses the repository's published storage only. The publisher's own runtime notes are the ones to follow.
Smallest pack we can size: Smallest published GGUF · 7.9 GB, including the allowance.
Qwen's 7B coder instruct model. Sized for a desktop that is editing code, not for a multi-GPU server.
Smallest pack we can size: Q4 estimate · 5.2 GB, including the allowance.
Qwen's 32B coder instruct model. Use the 4-bit row on a 24 GB card and the rented row if that estimate does not fit.
Smallest pack we can size: Q4 estimate · 20.5 GB, including the allowance.
DeepSeek Coder V2 Lite, instruct weights. Lite is the publisher's name. The gigabytes still come from the parameter count on the record.
Smallest pack we can size: Q4 estimate · 10.1 GB, including the allowance.
Black Forest Labs' FLUX.1 schnell, a text-to-image model built for few sampling steps. The fit row is weight memory plus a fixed activation allowance. It is not a render this desk timed.
Smallest pack we can size: FP8 estimate · 15.9 GB, including the allowance.
FLUX.1 dev from Black Forest Labs. The license is not the same as schnell. Read it before any commercial image work.
Smallest pack we can size: FP8 estimate · 15.9 GB, including the allowance.
A Comfy-Org package of the Flux 2 Klein 4B image model. The publisher of this repository is Comfy-Org. Follow their file layout if you run it in ComfyUI.
Smallest pack we can size: Largest published checkpoint · 11.8 GB, including the allowance.
Stability AI's Stable Diffusion XL base checkpoint. Widely used, and older than the Flux models listed here. The license row is the one to read.
Smallest pack we can size: FP8 estimate · 6.6 GB, including the allowance.
Wan-AI's 1.3B text-to-video model in Diffusers layout. Small for a video model. The activation allowance on a video row is larger than on a chat row, because frames are not tokens.
Smallest pack we can size: FP8 estimate · 9.4 GB, including the allowance.
The 14B Wan 2.1 text-to-video checkpoint in Diffusers layout. Expect a rented GPU. A 24 GB card is the borderline case on the FP8 estimate, not a promise.
Smallest pack we can size: FP8 estimate · 22.3 GB, including the allowance.
Lightricks' LTX-Video model. The public record classes it as image-to-video. Run it from the publisher's instructions. The fit number is an estimate.
Smallest pack we can size: FP8 estimate · 9.9 GB, including the allowance.
MiniMax H3, published for image, text, and video. The parameter count on a model this large usually means a rented multi-GPU machine. Read the community license before you use the outputs.
Smallest pack we can size: FP8 estimate · 41.1 GB, including the allowance.
OpenAI's Whisper large-v3 speech-to-text model. It is small next to a chat model. The usual question is language coverage, which the model card states and this page does not re-score.
Smallest pack we can size: FP16 estimate · 4.1 GB, including the allowance.
The faster Whisper large-v3 turbo checkpoint from OpenAI. Same job as large-v3: transcribe audio on a machine you control.
Smallest pack we can size: FP16 estimate · 2.1 GB, including the allowance.
Kokoro, an 82M text-to-speech model published by hexgrad. The weights are tiny. The voice files and the license live in the publisher repository.
Smallest pack we can size: Largest published checkpoint · 0.8 GB, including the allowance.
BAAI's bge-m3 embedding model for search. It turns text into vectors. It does not chat. The fit row should be an easy pass on almost any current machine.
Smallest pack we can size: Largest published checkpoint · 3.3 GB, including the allowance.
BAAI's bge reranker v2 m3. It reorders search hits. Pair it with an embedding model. It is not a chat model.
Smallest pack we can size: Q4 estimate · 1.3 GB, including the allowance.
This directory is assembled from public model records. Freshness is age. Popularity is downloads and likes. A fit figure is an estimate from a parameter count or from a published file size, not a run on our hardware. How the numbers are made. AI disclosure.